def prune_linear(O0O0OOO00OO000OOO: nn.modules.linear.Linear,
O0O0OOOO00O0O0OOO: list,
inplace: list = True,
dry_run: list = False): #line:85
"" #line:91
O0O0000OOO000OO00 = len(
O0O0OOOO00O0O0OOO) * O0O0OOO00OO000OOO.weight.shape[1] + (
len(O0O0OOOO00O0O0OOO) if O0O0OOO00OO000OOO.bias is not None else 0
) #line:92
if dry_run: #line:93
return O0O0OOO00OO000OOO, O0O0000OOO000OO00 #line:94
if not inplace: #line:96
O0O0OOO00OO000OOO = deepcopy(O0O0OOO00OO000OOO) #line:97
OOOO00O0OOOOO00OO = [
O0O0O000OO0000O0O
for O0O0O000OO0000O0O in range(O0O0OOO00OO000OOO.out_features)
if O0O0O000OO0000O0O not in O0O0OOOO00O0O0OOO
] #line:98
O0O0OOO00OO000OOO.out_features = O0O0OOO00OO000OOO.out_features - len(
O0O0OOOO00O0O0OOO) #line:99
O0O0OOO00OO000OOO.weight = torch.nn.Parameter(
O0O0OOO00OO000OOO.weight.data.clone()[OOOO00O0OOOOO00OO]) #line:100
if O0O0OOO00OO000OOO.bias is not None: #line:101
O0O0OOO00OO000OOO.bias = torch.nn.Parameter(
O0O0OOO00OO000OOO.bias.data.clone()[OOOO00O0OOOOO00OO]) #line:102
return O0O0OOO00OO000OOO, O0O0000OOO000OO00 #line:103
def prune_related_linear(layer: nn.modules.linear.Linear,
idxs: list,
inplace: list = True,
dry_run: list = False): #line:105
"" #line:111
O00O00O00O0000O0O = len(idxs) * layer.weight.shape[0] #line:112
if dry_run: #line:113
return layer, O00O00O00O0000O0O #line:114
if not inplace: #line:116
layer = deepcopy(layer) #line:117
OO0OOOO0O0O0000OO = [
O00OO0O000O0O0O0O for O00OO0O000O0O0O0O in range(layer.in_features)
if O00OO0O000O0O0O0O not in idxs
] #line:118
layer.in_features = layer.in_features - len(idxs) #line:119
layer.weight = torch.nn.Parameter(
layer.weight.data.clone()[:, OO0OOOO0O0O0000OO]) #line:120
return layer, O00O00O00O0000O0O #line:121
def prune_related_linear(layer: nn.modules.linear.Linear,
idxs: list,
inplace: list = True,
dry_run: list = False):
"""Prune weights for the related (affected) fully-connected layer, e.g. [256 x 128] => [256 x 96]
Args:
layer: a fully-connected layer.
idxs: pruning index.
"""
num_pruned = len(idxs) * layer.weight.shape[0]
if dry_run:
return layer, num_pruned
if not inplace:
layer = deepcopy(layer)
keep_idxs = [i for i in range(layer.in_features) if i not in idxs]
layer.in_features = layer.in_features - len(idxs)
layer.weight = torch.nn.Parameter(layer.weight.data.clone()[:, keep_idxs])
return layer, num_pruned
def prune_related_linear(OOO00000O0OOO0000: nn.modules.linear.Linear,
O00OOOOOOOOOO000O: list,
inplace: list = True,
dry_run: list = False): #line:105
"" #line:111
OO0OOO00O000O00O0 = len(
O00OOOOOOOOOO000O) * OOO00000O0OOO0000.weight.shape[0] #line:112
if dry_run: #line:113
return OOO00000O0OOO0000, OO0OOO00O000O00O0 #line:114
if not inplace: #line:116
OOO00000O0OOO0000 = deepcopy(OOO00000O0OOO0000) #line:117
O0O00000OOO0O0OO0 = [
OOO0O000O00OOO00O
for OOO0O000O00OOO00O in range(OOO00000O0OOO0000.in_features)
if OOO0O000O00OOO00O not in O00OOOOOOOOOO000O
] #line:118
OOO00000O0OOO0000.in_features = OOO00000O0OOO0000.in_features - len(
O00OOOOOOOOOO000O) #line:119
OOO00000O0OOO0000.weight = torch.nn.Parameter(
OOO00000O0OOO0000.weight.data.clone()[:, O0O00000OOO0O0OO0]) #line:120
return OOO00000O0OOO0000, OO0OOO00O000O00O0 #line:121
def prune_linear(layer: nn.modules.linear.Linear,
idxs: list,
inplace: list = True,
dry_run: list = False): #line:85
"" #line:91
OOO00OO00O000OOO0 = len(idxs) * layer.weight.shape[1] + (
len(idxs) if layer.bias is not None else 0) #line:92
if dry_run: #line:93
return layer, OOO00OO00O000OOO0 #line:94
if not inplace: #line:96
layer = deepcopy(layer) #line:97
O0OOOOOOO000000OO = [
OOO0OO000OO0000O0 for OOO0OO000OO0000O0 in range(layer.out_features)
if OOO0OO000OO0000O0 not in idxs
] #line:98
layer.out_features = layer.out_features - len(idxs) #line:99
layer.weight = torch.nn.Parameter(
layer.weight.data.clone()[O0OOOOOOO000000OO]) #line:100
if layer.bias is not None: #line:101
layer.bias = torch.nn.Parameter(
layer.bias.data.clone()[O0OOOOOOO000000OO]) #line:102
return layer, OOO00OO00O000OOO0 #line:103
def prune_linear(layer: nn.modules.linear.Linear,
idxs: list,
inplace: list = True,
dry_run: list = False):
"""Prune neurons for the fully-connected layer, e.g. [256 x 128] => [192 x 128]
Args:
layer: a fully-connected layer.
idxs: pruning index.
"""
num_pruned = len(idxs) * layer.weight.shape[1] + (len(idxs) if layer.bias
is not None else 0)
if dry_run:
return layer, num_pruned
if not inplace:
layer = deepcopy(layer)
keep_idxs = [i for i in range(layer.out_features) if i not in idxs]
layer.out_features = layer.out_features - len(idxs)
layer.weight = torch.nn.Parameter(layer.weight.data.clone()[keep_idxs])
if layer.bias is not None:
layer.bias = torch.nn.Parameter(layer.bias.data.clone()[keep_idxs])
return layer, num_pruned